| --- |
| license: mit |
| task_categories: |
| - other |
| tags: |
| - mechanistic-interpretability |
| - robustness |
| - language-model-evaluation |
| - perturbation-analysis |
| --- |
| |
| # Decoding Robustness Results |
|
|
| Mechanistic robustness evaluation results for language models under six input |
| perturbations: character replacement, BPE-token replacement, word replacement, |
| local token shuffle, typographical corruption, and synonym replacement. |
|
|
| The repository is organized by model and perturbation: |
|
|
| ```text |
| models/<model>/<perturbation>/<percentage>/evals.csv |
| ``` |
|
|
| The `qwen2.5_1.5b/adversarial` directory contains the separate adversarial |
| evaluation outputs and manifest. Failed or mislabeled Kaggle runs are not |
| included in the canonical result directories. |
|
|
| Recovered model coverage currently includes complete sweeps for GPT-2, |
| GPT-2 Medium, GPT-2 XL, Qwen 2.5 0.5B, and Qwen 2.5 1.5B. GPT-2 Large is |
| represented by a partial legacy run under `models/gpt2-large/`; it contains |
| 11 `char` percentages and 2 valid `token` percentages. An empty GPT-2 Large |
| `token/20` file and an empty legacy GPT-2 XL file were excluded as invalid |
| results. The GPT-2 XL split-A and split-B outputs are merged under |
| `models/gpt2-xl/`. |
|
|
| Source dataset: WikiText-2 raw test split (`Salesforce/wikitext`, |
| `wikitext-2-raw-v1`). |
|
|
